Feature Extraction for Analysis of Electron MicroscopyImages

Andrew Kennardakennard@stanford.edu

October 30, 2015

Transmission electron microscopy (TEM) provides extremely detailed high-resolution im-ages of biological structures ranging from subcellular organelles to whole tissues. An ad-vantage of this approach is that TEM allows simultaneous visualization of many differentcellular components, providing a holistic view of a cell or tissue, in contrast to fluorescencemicroscopy which provides spatial information about only a few different types of cellularcomponents in any one experiment.

However, the wealth of information in such images makes it diffult to extract particulardetails from many images at once. The most common type of information desired fromelectron microscopy images is quantitative morphology of cellular components, such as thenumber, shape, and distribution of particular organelles within a cell or the number andtopology of connections between a group of cells.

A further challenge of electron microscopy data analysis is that in order to see whole cellsat the fine resolution afforded by TEM requires tiling from many smaller images which mustbe computationally stitched together to create a larger image (which is then of considerablesize). A final challenge of TEM is registering different slices of a cell to create a 3D recon-struction of a cell at high resolution. Due to the number of different processing steps requiredto create a complete EM reconstruction of a cellular volume, automated or semi-automatedalgorithms for feature extraction, image stitching, and tomography are crucial for efficientelectron microscopy analysis.

Useful semi-automated tools are available for stitching together images and creating 3Dtomograms Cardona et al. [2012], but this software is very general and hence requires signif-icant manual input for any particular task, and furthermore does not provide many tools forsophisticated detection of relevant features in EM images. For the dedicated electron micro-scopist there is instead a need for more specialized tools optimized to streamline analysis ofspecific types of images, in particular the detection of particular features of interest.

1 Specific proposal

I propose to develop such a tool to specifically automate the process of imaging of a particulartype of cell: a skin cell from fish (Fig. 1), which has a very dense meshwork of protein

1

Kennard, Andrew EE368/CS232 Final Project Proposal 29 Oct. 2015

Figure 1: Electron micrograph of a fish skin cell. There are many different textures inthis image representing different cellular features. I propose to develop an algorithm todistinguish some particular textures of interest.

filaments called the actin cytoskeleton that is of particular fascination to biologists interestedin how cells move or transduce mechanical forces from their environment.

I will develop an image processing algorithm that will specifically detect this dense mesh-work in a series of EM images and segment out this feature. I will then be able to feed thisdata into existing algorithms to create 3D reconstructions of the location of this meshworkfeature for high-resolution analysis of cellular structure. This algorithm would be highlyrelevant to my research.

2 Techniques and Approaches

I expect that successfully identifying my features of interest will require a combination ofdifferent processing approaches. One approach will be to exploit the dark intensity of thisarea to try simple smoothing followed by thresholding of the image. However, there aremany other dark regions in the image and I expect I will find many false positives.

To improve this approach, I will also try implementing feature detection strategies thatrely on texture analysis, looking for more consistent or repetitive textures in images. Onetechnique that has been successfully used on EM images of viruses is the local co-occurrencematrix of gray levels Proenca et al. [2013], which looks for spatial correlations in the grayscaleintensities of an image. Another technique that I will try is the use of Gabor filters, whichare sensitive to orientational information and may be able to detect certain types of texturesin these images well Gorai et al. [2014].

I will NOT be using Android devices for my project.

2

Kennard, Andrew EE368/CS232 Final Project Proposal 29 Oct. 2015

References

Albert Cardona, Stephan Saalfeld, Johannes Schindelin, Ignacio Arganda-Carreras,Stephan Preibisch, Mark Longair, Pavel Tomancak, Volker Hartenstein, and Rod-ney J. Douglas. TrakEM2 Software for Neural Circuit Reconstruction. PLoS ONE,7(6):e38011, 2012. ISSN 1932-6203. doi: 10.1371/journal.pone.0038011. URLhttp://dx.plos.org/10.1371/journal.pone.0038011.

Apurba Gorai, Kendrick Cetina, and Luis Baumela. A comparative study of local binarypattern descriptors and Gabor filter for electron microscopy image segmentation. Inter-national Conference on Parallel, Distributed, and Grid Computing, pages 76–81, 2014.

Maria C Proenca, Jose F M Nunes, and Antonio P A de Matos. Texture indica-tors for segmentation of polyomavirus particles in transmission electron microscopyimages. Microscopy and microanalysis, 19(5):1170–82, 2013. ISSN 1435-8115. doi:10.1017/S1431927613001736. URL http://www.ncbi.nlm.nih.gov/pubmed/23773502.

3